IEEE Transactions on Evolutionary Computation, 2018 ESI Highly Cited Paper

A Classification-Based Surrogate-Assisted Evolutionary Algorithm for Expensive Many-Objective Optimization

Linqiang Pan, Cheng He, Ye Tian, Handing Wang, Xingyi Zhang, Yaochu Jin*

Abstract

Surrogate-assisted evolutionary algorithms have been widely developed for expensive optimization problems where each function evaluation incurs substantial computational cost. However, extending surrogates to many-objective optimization (4+ objectives) presents unique challenges: the surrogate must accurately predict not only objective values but also dominance relationships in high-dimensional objective space. This work introduces CSEA, which reframes surrogate modeling as a classification problem. Rather than regressing individual objective values, a classifier is trained to predict whether a candidate solution dominates a reference set of solutions — a binary decision that is substantially easier to learn and more robust. A feedforward neural network serves as the classifier, with training data generated from archived non-dominated solutions. CSEA demonstrated state-of-the-art performance on expensive many-objective benchmarks (up to 10 objectives), establishing classification-based surrogates as an effective paradigm for expensive evolutionary optimization.

Paper figures

CSEA: Classification-based surrogate-assisted evolutionary algorithm framework

Paper figure: Algorithm framework diagram of CSEA. The figure illustrates how the neural network classifier is trained on archived non-dominated solutions to predict dominance relationships, replacing expensive regression-based surrogates with a more robust classification task. This design enables efficient many-objective optimization with limited function evaluations.

Key contributions

  • Classification-based surrogate modeling reframes the surrogate task from regression (predicting objective values) to classification (predicting dominance), which is more robust in high-dimensional objective space.
  • A neural network classifier trained on archived non-dominated solutions predicts whether candidate solutions are promising, dramatically reducing expensive function evaluations.
  • Validated on benchmarks with up to 10 objectives and real-world engineering problems, establishing CSEA as one of the most-cited surrogate-assisted many-objective EAs.